From: Ian Goldberg <iang@CS.Berkeley.EDU>
Date: Sun, 17 Sep 95 21:41:08 PDT
To: cypherpunks@toad.com
Subject: Netscape SSL implementation cracked!
Message-ID: <199509180441.VAA16683@lagos.CS.Berkeley.EDU>
MIME-Version: 1.0
Content-Type: text/plain


As some of you may recall, a few weeks ago I posted a
reverse-compilation of the random number generation routine used by
netscape to choose challenge data and encryption keys.

Recently, one of my officemates (David Wagner <daw@cs.berkeley.edu>)
and I (Ian Goldberg <iang@cs.berkeley.edu>) finished the job
of seeing exactly how the encryption keys are picked.

What we discovered is that, at least on the systems we checked (Solaris
and HP-UX), the seed value for the RNG was fairly trivial to guess by
someone with an account on the machine running netscape (so much so
that in this situation, it usually takes less than 1 minute to find
the key), and not too hard for people without accounts, either.
See below for details.

I've included the header to a program we wrote to do this key-cracking
below.  I would like to get some information, though:

o Where should I put the full source (1 file, ~12k) so that ITAR lovers
  don't get mad at me?

o Where can I find a version of netscape that does RC4-128?  It is
  likely that it suffers from the same problem, and even a brute-force
  search of the entire seed space is _much_ less than 128 bits.

Happy hacking,

   - Ian "who just saw _Hackers_ today with some other Bay Area cypherpunks,
           and it put me in the mood"

/* unssl.c - Last update: 950917

   Break netscape's shoddy implementation of SSL on some platforms
   (tested for netscape running RC4-40 on Solaris and HP-UX; other
    Unices are probably similar; other crypt methods are unknown, but
    it is likely that RC4-128 will have the same problems).

   The idea is this: netscape seeds the random number generator it uses
   to produce challenge-data and master keys with a combination of the
   time in seconds and microseconds, the pid and the ppid.  Of these,
   only the microseconds is hard to determine by someone who
   (a) can watch your packets on the network and
   (b) has access to any account on the system running netscape.

   Even if (b) is not satisfied, the time can often be obtained from
   the time or daytime network daemons; an approximation to the pid can
   sometimes be obtained from a mail daemon (the pid is part of most
   Message-ID's); the ppid will usually be not much smaller than the pid,
   and has an higher than average chance of being 1.  Clever guessing
   of these values will in all likelihood cut the expected search space
   down to less than brute-forcing a 40-bit key, and certainly is less
   than brute-forcing a 128-bit key.

   Subsequent https: connections after the first (even to different hosts)
   seem to _not_ reseed the RNG.  This makes things much easier, once
   you've broken the first message.  Just keep generating 16 bytes of
   random numbers until you get the challenge-data for the next message.
   The next key will then be the 16 random bytes after that.

   main() and bits of MD5Transform1 by Ian Goldberg <iang@cs.berkeley.edu>
   and David Wagner <daw@cs.berkeley.edu>.  The rest is taken from the
   standard MD5 code; see below.

   This code seems to want to run on a big-endian machine.  There may be
   other problems as well.  This code is provided as-is; if it causes you
   to lose your data, sleep, civil liberties, or SO, that's your problem.

#include <std/disclaimer.h>

   On the command line, give the time in seconds, the pid, the ppid and
   the SSL challenge data (each byte in hex, separated by some non-hex
   character like a colon) of the _first_ SSL message generated by
   the instance of netscape.  This program will search through the
   microsecond values.  You may need to run it again with a slightly
   different value for the seconds, depending on how accurately you know
   the time on the system running netscape.  The output will be the
   master key (all 16 bytes; note you never even told the program the
   11 bytes you knew) and the value for the microseconds that produced it.

   As a benchmark, this code runs in just under 25 seconds real time
   (for an unsuccessful search through 1<<20 values for the microseconds)
   on an unloaded HP 712/80.
*/